Optimizing deep neural networks for high-resolution land cover classification through data augmentation

• DOI: 10.1007/s10661-025-13870-5
• Publication Date: 2025-03-18T16:53:56Z
• Authors: Sergio Sierra; Rubén Ramo; Marc Padilla; Adolfo Cobo
• Keywords: Land Cover

Abstract This study presents an innovative approach to high-resolution land cover classification using deep learning, tackling the challenge of working with exceptionally small dataset. Manual annotation data is both time-consuming and labor-intensive, making augmentation crucial for enhancing model performance. While a well-established technique, there has not been comprehensive comparative evaluation wide range methods specifically applied until now. Our work fills this gap by systematically testing eight different techniques across four neural networks (U-Net, DeepLabv3 + , FCN, PSPNet) 25 cm resolution images from Cantabria, Spain. In total, we generated 19 distinct training sets trained validated 72 models. The results show that can boost performance up 30%. best (DeepLabV3 flip, contrast, brightness adjustments) achieved accuracy 0.89 IoU 0.78. Additionally, utilized optimized generate maps years 2014, 2017, 2019, at 580 samples selected based on stratified sampling CORINE Land Cover data, achieving 87.2%. only provides systematic ranking but also offers practical framework help future researchers save time identifying most effective strategies specific task.